Useful agents need bounded capabilities
An agentic commerce system lets an AI assistant use application capabilities to help with a commerce task. The important design choice is the boundary around those capabilities: what the agent can read, what it can propose, and what requires a person's authorization.
A conversational interface alone does not establish that boundary. The application must enforce it, independently of whatever the model says.
Ground answers in commerce facts
Product discovery requires authoritative product identifiers, current prices and availability. Knowledge retrieval can explain a policy or a buying consideration, but it should not substitute for current catalog facts.
Rocento's assistant preview uses scoped read-only tools through a commerce facade. Tenant and persona boundaries limit the available context. Approved knowledge retrieval is a separate source with its own provenance.
Separate reasoning from permission
A model may propose a next step. That proposal is not authorization. The application must validate parameters and scope before carrying out any permitted capability.
Customer request
↓
Agent reasoning
↓
Approved tool + scoped arguments
↓
Application authorization and validation
↓
Authoritative result
This conceptual sequence deliberately avoids a direct connection between a model and an unrestricted database. It also keeps commerce invariants in the application where ordinary user actions are validated.
Handle incomplete answers honestly
A product search can return no matches. A service can be unavailable. A policy source can be missing. These cases should produce a clear limitation or a clarifying question, rather than fabricated products or confident guesses.
For example, a request for a cheaper alternative needs a known reference product and a fresh price. If either is missing, asking for clarification is a useful outcome.
The current scope
The Rocento assistant preview includes catalog-grounded shopping assistance and documented order-status and support flows. The agent commerce tools described here are read-only; this does not imply autonomous purchasing, inventory changes or merchant operations. Broader autonomous actions remain in development.
Production use requires evaluation of answer quality, provider configuration, access controls and operational readiness for the intended deployment.
Evaluate the boundary as well as the answer
Future evaluation should ask whether the system respects access scope, handles missing facts, preserves explicit customer constraints and accurately reports tool failures. Fluent wording is only one aspect of a useful assistant.